drivers/net/xen-netback/netback.c

   1 /*
   2  * Back-end of the driver for virtual network devices. This portion of the
   3  * driver exports a 'unified' network-device interface that can be accessed
   4  * by any operating system that implements a compatible front end. A
   5  * reference front-end implementation can be found in:
   6  *  drivers/net/xen-netfront.c
   7  *
   8  * Copyright (c) 2002-2005, K A Fraser
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation; or, when distributed
  13  * separately from the Linux kernel or incorporated into other
  14  * software packages, subject to the following license:
  15  *
  16  * Permission is hereby granted, free of charge, to any person obtaining a copy
  17  * of this source file (the "Software"), to deal in the Software without
  18  * restriction, including without limitation the rights to use, copy, modify,
  19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  20  * and to permit persons to whom the Software is furnished to do so, subject to
  21  * the following conditions:
  22  *
  23  * The above copyright notice and this permission notice shall be included in
  24  * all copies or substantial portions of the Software.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  32  * IN THE SOFTWARE.
  33  */
  34
  35 #include "common.h"
  36
  37 #include <linux/kthread.h>
  38 #include <linux/if_vlan.h>
  39 #include <linux/udp.h>
  40 #include <linux/highmem.h>
  41 #include <linux/skbuff_ref.h>
  42
  43 #include <net/tcp.h>
  44
  45 #include <xen/xen.h>
  46 #include <xen/events.h>
  47 #include <xen/interface/memory.h>
  48 #include <xen/page.h>
  49
  50 #include <asm/xen/hypercall.h>
  51
  52 /* Provide an option to disable split event channels at load time as
  53  * event channels are limited resource. Split event channels are
  54  * enabled by default.
  55  */
  56 bool separate_tx_rx_irq = true;
  57 module_param(separate_tx_rx_irq, bool, 0644);
  58
  59 /* The time that packets can stay on the guest Rx internal queue
  60  * before they are dropped.
  61  */
  62 unsigned int rx_drain_timeout_msecs = 10000;
  63 module_param(rx_drain_timeout_msecs, uint, 0444);
  64
  65 /* The length of time before the frontend is considered unresponsive
  66  * because it isn't providing Rx slots.
  67  */
  68 unsigned int rx_stall_timeout_msecs = 60000;
  69 module_param(rx_stall_timeout_msecs, uint, 0444);
  70
  71 #define MAX_QUEUES_DEFAULT 8
  72 unsigned int xenvif_max_queues;
  73 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  74 MODULE_PARM_DESC(max_queues,
  75                  "Maximum number of queues per virtual interface");
  76
  77 /*
  78  * This is the maximum slots a skb can have. If a guest sends a skb
  79  * which exceeds this limit it is considered malicious.
  80  */
  81 #define FATAL_SKB_SLOTS_DEFAULT 20
  82 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  83 module_param(fatal_skb_slots, uint, 0444);
  84
  85 /* The amount to copy out of the first guest Tx slot into the skb's
  86  * linear area.  If the first slot has more data, it will be mapped
  87  * and put into the first frag.
  88  *
  89  * This is sized to avoid pulling headers from the frags for most
  90  * TCP/IP packets.
  91  */
  92 #define XEN_NETBACK_TX_COPY_LEN 128
  93
  94 /* This is the maximum number of flows in the hash cache. */
  95 #define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
  96 unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
  97 module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
  98 MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
  99
 100 /* The module parameter tells that we have to put data
 101  * for xen-netfront with the XDP_PACKET_HEADROOM offset
 102  * needed for XDP processing
 103  */
 104 bool provides_xdp_headroom = true;
 105 module_param(provides_xdp_headroom, bool, 0644);
 106
 107 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
 108                                s8 status);
 109
 110 static void make_tx_response(struct xenvif_queue *queue,
 111                              const struct xen_netif_tx_request *txp,
 112                              unsigned int extra_count,
 113                              s8 status);
 114
 115 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
 116
 117 static inline int tx_work_todo(struct xenvif_queue *queue);
 118
 119 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 120                                        u16 idx)
 121 {
 122         return page_to_pfn(queue->mmap_pages[idx]);
 123 }
 124
 125 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 126                                          u16 idx)
 127 {
 128         return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 129 }
 130
 131 #define callback_param(vif, pending_idx) \
 132         (vif->pending_tx_info[pending_idx].callback_struct)
 133
 134 /* Find the containing VIF's structure from a pointer in pending_tx_info array
 135  */
 136 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
 137 {
 138         u16 pending_idx = ubuf->desc;
 139         struct pending_tx_info *temp =
 140                 container_of(ubuf, struct pending_tx_info, callback_struct);
 141         return container_of(temp - pending_idx,
 142                             struct xenvif_queue,
 143                             pending_tx_info[0]);
 144 }
 145
 146 static u16 frag_get_pending_idx(skb_frag_t *frag)
 147 {
 148         return (u16)skb_frag_off(frag);
 149 }
 150
 151 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 152 {
 153         skb_frag_off_set(frag, pending_idx);
 154 }
 155
 156 static inline pending_ring_idx_t pending_index(unsigned i)
 157 {
 158         return i & (MAX_PENDING_REQS-1);
 159 }
 160
 161 void xenvif_kick_thread(struct xenvif_queue *queue)
 162 {
 163         wake_up(&queue->wq);
 164 }
 165
 166 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 167 {
 168         int more_to_do;
 169
 170         RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 171
 172         if (more_to_do)
 173                 napi_schedule(&queue->napi);
 174         else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 175                                      &queue->eoi_pending) &
 176                  (NETBK_TX_EOI | NETBK_COMMON_EOI))
 177                 xen_irq_lateeoi(queue->tx_irq, 0);
 178 }
 179
 180 static void tx_add_credit(struct xenvif_queue *queue)
 181 {
 182         unsigned long max_burst, max_credit;
 183
 184         /*
 185          * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 186          * Otherwise the interface can seize up due to insufficient credit.
 187          */
 188         max_burst = max(131072UL, queue->credit_bytes);
 189
 190         /* Take care that adding a new chunk of credit doesn't wrap to zero. */
 191         max_credit = queue->remaining_credit + queue->credit_bytes;
 192         if (max_credit < queue->remaining_credit)
 193                 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 194
 195         queue->remaining_credit = min(max_credit, max_burst);
 196         queue->rate_limited = false;
 197 }
 198
 199 void xenvif_tx_credit_callback(struct timer_list *t)
 200 {
 201         struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
 202         tx_add_credit(queue);
 203         xenvif_napi_schedule_or_enable_events(queue);
 204 }
 205
 206 static void xenvif_tx_err(struct xenvif_queue *queue,
 207                           struct xen_netif_tx_request *txp,
 208                           unsigned int extra_count, RING_IDX end)
 209 {
 210         RING_IDX cons = queue->tx.req_cons;
 211
 212         do {
 213                 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 214                 if (cons == end)
 215                         break;
 216                 RING_COPY_REQUEST(&queue->tx, cons++, txp);
 217                 extra_count = 0; /* only the first frag can have extras */
 218         } while (1);
 219         queue->tx.req_cons = cons;
 220 }
 221
 222 static void xenvif_fatal_tx_err(struct xenvif *vif)
 223 {
 224         netdev_err(vif->dev, "fatal error; disabling device\n");
 225         vif->disabled = true;
 226         /* Disable the vif from queue 0's kthread */
 227         if (vif->num_queues)
 228                 xenvif_kick_thread(&vif->queues[0]);
 229 }
 230
 231 static int xenvif_count_requests(struct xenvif_queue *queue,
 232                                  struct xen_netif_tx_request *first,
 233                                  unsigned int extra_count,
 234                                  struct xen_netif_tx_request *txp,
 235                                  int work_to_do)
 236 {
 237         RING_IDX cons = queue->tx.req_cons;
 238         int slots = 0;
 239         int drop_err = 0;
 240         int more_data;
 241
 242         if (!(first->flags & XEN_NETTXF_more_data))
 243                 return 0;
 244
 245         do {
 246                 struct xen_netif_tx_request dropped_tx = { 0 };
 247
 248                 if (slots >= work_to_do) {
 249                         netdev_err(queue->vif->dev,
 250                                    "Asked for %d slots but exceeds this limit\n",
 251                                    work_to_do);
 252                         xenvif_fatal_tx_err(queue->vif);
 253                         return -ENODATA;
 254                 }
 255
 256                 /* This guest is really using too many slots and
 257                  * considered malicious.
 258                  */
 259                 if (unlikely(slots >= fatal_skb_slots)) {
 260                         netdev_err(queue->vif->dev,
 261                                    "Malicious frontend using %d slots, threshold %u\n",
 262                                    slots, fatal_skb_slots);
 263                         xenvif_fatal_tx_err(queue->vif);
 264                         return -E2BIG;
 265                 }
 266
 267                 /* Xen network protocol had implicit dependency on
 268                  * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 269                  * the historical MAX_SKB_FRAGS value 18 to honor the
 270                  * same behavior as before. Any packet using more than
 271                  * 18 slots but less than fatal_skb_slots slots is
 272                  * dropped
 273                  */
 274                 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 275                         if (net_ratelimit())
 276                                 netdev_dbg(queue->vif->dev,
 277                                            "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 278                                            slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 279                         drop_err = -E2BIG;
 280                 }
 281
 282                 if (drop_err)
 283                         txp = &dropped_tx;
 284
 285                 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 286
 287                 /* If the guest submitted a frame >= 64 KiB then
 288                  * first->size overflowed and following slots will
 289                  * appear to be larger than the frame.
 290                  *
 291                  * This cannot be fatal error as there are buggy
 292                  * frontends that do this.
 293                  *
 294                  * Consume all slots and drop the packet.
 295                  */
 296                 if (!drop_err && txp->size > first->size) {
 297                         if (net_ratelimit())
 298                                 netdev_dbg(queue->vif->dev,
 299                                            "Invalid tx request, slot size %u > remaining size %u\n",
 300                                            txp->size, first->size);
 301                         drop_err = -EIO;
 302                 }
 303
 304                 first->size -= txp->size;
 305                 slots++;
 306
 307                 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 308                         netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 309                                  txp->offset, txp->size);
 310                         xenvif_fatal_tx_err(queue->vif);
 311                         return -EINVAL;
 312                 }
 313
 314                 more_data = txp->flags & XEN_NETTXF_more_data;
 315
 316                 if (!drop_err)
 317                         txp++;
 318
 319         } while (more_data);
 320
 321         if (drop_err) {
 322                 xenvif_tx_err(queue, first, extra_count, cons + slots);
 323                 return drop_err;
 324         }
 325
 326         return slots;
 327 }
 328
 329
 330 struct xenvif_tx_cb {
 331         u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
 332         u8 copy_count;
 333         u32 split_mask;
 334 };
 335
 336 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 337 #define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
 338 #define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
 339
 340 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 341                                            u16 pending_idx,
 342                                            struct xen_netif_tx_request *txp,
 343                                            unsigned int extra_count,
 344                                            struct gnttab_map_grant_ref *mop)
 345 {
 346         queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 347         gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 348                           GNTMAP_host_map | GNTMAP_readonly,
 349                           txp->gref, queue->vif->domid);
 350
 351         memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 352                sizeof(*txp));
 353         queue->pending_tx_info[pending_idx].extra_count = extra_count;
 354 }
 355
 356 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 357 {
 358         struct sk_buff *skb =
 359                 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 360                           GFP_ATOMIC | __GFP_NOWARN);
 361
 362         BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
 363         if (unlikely(skb == NULL))
 364                 return NULL;
 365
 366         /* Packets passed to netif_rx() must have some headroom. */
 367         skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 368
 369         /* Initialize it here to avoid later surprises */
 370         skb_shinfo(skb)->destructor_arg = NULL;
 371
 372         return skb;
 373 }
 374
 375 static void xenvif_get_requests(struct xenvif_queue *queue,
 376                                 struct sk_buff *skb,
 377                                 struct xen_netif_tx_request *first,
 378                                 struct xen_netif_tx_request *txfrags,
 379                                 unsigned *copy_ops,
 380                                 unsigned *map_ops,
 381                                 unsigned int frag_overflow,
 382                                 struct sk_buff *nskb,
 383                                 unsigned int extra_count,
 384                                 unsigned int data_len)
 385 {
 386         struct skb_shared_info *shinfo = skb_shinfo(skb);
 387         skb_frag_t *frags = shinfo->frags;
 388         u16 pending_idx;
 389         pending_ring_idx_t index;
 390         unsigned int nr_slots;
 391         struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
 392         struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
 393         struct xen_netif_tx_request *txp = first;
 394
 395         nr_slots = shinfo->nr_frags + frag_overflow + 1;
 396
 397         copy_count(skb) = 0;
 398         XENVIF_TX_CB(skb)->split_mask = 0;
 399
 400         /* Create copy ops for exactly data_len bytes into the skb head. */
 401         __skb_put(skb, data_len);
 402         while (data_len > 0) {
 403                 int amount = data_len > txp->size ? txp->size : data_len;
 404                 bool split = false;
 405
 406                 cop->source.u.ref = txp->gref;
 407                 cop->source.domid = queue->vif->domid;
 408                 cop->source.offset = txp->offset;
 409
 410                 cop->dest.domid = DOMID_SELF;
 411                 cop->dest.offset = (offset_in_page(skb->data +
 412                                                    skb_headlen(skb) -
 413                                                    data_len)) & ~XEN_PAGE_MASK;
 414                 cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
 415                                                - data_len);
 416
 417                 /* Don't cross local page boundary! */
 418                 if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
 419                         amount = XEN_PAGE_SIZE - cop->dest.offset;
 420                         XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
 421                         split = true;
 422                 }
 423
 424                 cop->len = amount;
 425                 cop->flags = GNTCOPY_source_gref;
 426
 427                 index = pending_index(queue->pending_cons);
 428                 pending_idx = queue->pending_ring[index];
 429                 callback_param(queue, pending_idx).ctx = NULL;
 430                 copy_pending_idx(skb, copy_count(skb)) = pending_idx;
 431                 if (!split)
 432                         copy_count(skb)++;
 433
 434                 cop++;
 435                 data_len -= amount;
 436
 437                 if (amount == txp->size) {
 438                         /* The copy op covered the full tx_request */
 439
 440                         memcpy(&queue->pending_tx_info[pending_idx].req,
 441                                txp, sizeof(*txp));
 442                         queue->pending_tx_info[pending_idx].extra_count =
 443                                 (txp == first) ? extra_count : 0;
 444
 445                         if (txp == first)
 446                                 txp = txfrags;
 447                         else
 448                                 txp++;
 449                         queue->pending_cons++;
 450                         nr_slots--;
 451                 } else {
 452                         /* The copy op partially covered the tx_request.
 453                          * The remainder will be mapped or copied in the next
 454                          * iteration.
 455                          */
 456                         txp->offset += amount;
 457                         txp->size -= amount;
 458                 }
 459         }
 460
 461         for (shinfo->nr_frags = 0; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
 462              nr_slots--) {
 463                 if (unlikely(!txp->size)) {
 464                         make_tx_response(queue, txp, 0, XEN_NETIF_RSP_OKAY);
 465                         ++txp;
 466                         continue;
 467                 }
 468
 469                 index = pending_index(queue->pending_cons++);
 470                 pending_idx = queue->pending_ring[index];
 471                 xenvif_tx_create_map_op(queue, pending_idx, txp,
 472                                         txp == first ? extra_count : 0, gop);
 473                 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 474                 ++shinfo->nr_frags;
 475                 ++gop;
 476
 477                 if (txp == first)
 478                         txp = txfrags;
 479                 else
 480                         txp++;
 481         }
 482
 483         if (nr_slots > 0) {
 484
 485                 shinfo = skb_shinfo(nskb);
 486                 frags = shinfo->frags;
 487
 488                 for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots; ++txp) {
 489                         if (unlikely(!txp->size)) {
 490                                 make_tx_response(queue, txp, 0,
 491                                                  XEN_NETIF_RSP_OKAY);
 492                                 continue;
 493                         }
 494
 495                         index = pending_index(queue->pending_cons++);
 496                         pending_idx = queue->pending_ring[index];
 497                         xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 498                                                 gop);
 499                         frag_set_pending_idx(&frags[shinfo->nr_frags],
 500                                              pending_idx);
 501                         ++shinfo->nr_frags;
 502                         ++gop;
 503                 }
 504
 505                 if (shinfo->nr_frags) {
 506                         skb_shinfo(skb)->frag_list = nskb;
 507                         nskb = NULL;
 508                 }
 509         }
 510
 511         if (nskb) {
 512                 /* A frag_list skb was allocated but it is no longer needed
 513                  * because enough slots were converted to copy ops above or some
 514                  * were empty.
 515                  */
 516                 kfree_skb(nskb);
 517         }
 518
 519         (*copy_ops) = cop - queue->tx_copy_ops;
 520         (*map_ops) = gop - queue->tx_map_ops;
 521 }
 522
 523 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 524                                            u16 pending_idx,
 525                                            grant_handle_t handle)
 526 {
 527         if (unlikely(queue->grant_tx_handle[pending_idx] !=
 528                      NETBACK_INVALID_HANDLE)) {
 529                 netdev_err(queue->vif->dev,
 530                            "Trying to overwrite active handle! pending_idx: 0x%x\n",
 531                            pending_idx);
 532                 BUG();
 533         }
 534         queue->grant_tx_handle[pending_idx] = handle;
 535 }
 536
 537 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 538                                              u16 pending_idx)
 539 {
 540         if (unlikely(queue->grant_tx_handle[pending_idx] ==
 541                      NETBACK_INVALID_HANDLE)) {
 542                 netdev_err(queue->vif->dev,
 543                            "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 544                            pending_idx);
 545                 BUG();
 546         }
 547         queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 548 }
 549
 550 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 551                                struct sk_buff *skb,
 552                                struct gnttab_map_grant_ref **gopp_map,
 553                                struct gnttab_copy **gopp_copy)
 554 {
 555         struct gnttab_map_grant_ref *gop_map = *gopp_map;
 556         u16 pending_idx;
 557         /* This always points to the shinfo of the skb being checked, which
 558          * could be either the first or the one on the frag_list
 559          */
 560         struct skb_shared_info *shinfo = skb_shinfo(skb);
 561         /* If this is non-NULL, we are currently checking the frag_list skb, and
 562          * this points to the shinfo of the first one
 563          */
 564         struct skb_shared_info *first_shinfo = NULL;
 565         int nr_frags = shinfo->nr_frags;
 566         const bool sharedslot = nr_frags &&
 567                                 frag_get_pending_idx(&shinfo->frags[0]) ==
 568                                     copy_pending_idx(skb, copy_count(skb) - 1);
 569         int i, err = 0;
 570
 571         for (i = 0; i < copy_count(skb); i++) {
 572                 int newerr;
 573
 574                 /* Check status of header. */
 575                 pending_idx = copy_pending_idx(skb, i);
 576
 577                 newerr = (*gopp_copy)->status;
 578
 579                 /* Split copies need to be handled together. */
 580                 if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
 581                         (*gopp_copy)++;
 582                         if (!newerr)
 583                                 newerr = (*gopp_copy)->status;
 584                 }
 585                 if (likely(!newerr)) {
 586                         /* The first frag might still have this slot mapped */
 587                         if (i < copy_count(skb) - 1 || !sharedslot)
 588                                 xenvif_idx_release(queue, pending_idx,
 589                                                    XEN_NETIF_RSP_OKAY);
 590                 } else {
 591                         err = newerr;
 592                         if (net_ratelimit())
 593                                 netdev_dbg(queue->vif->dev,
 594                                            "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 595                                            (*gopp_copy)->status,
 596                                            pending_idx,
 597                                            (*gopp_copy)->source.u.ref);
 598                         /* The first frag might still have this slot mapped */
 599                         if (i < copy_count(skb) - 1 || !sharedslot)
 600                                 xenvif_idx_release(queue, pending_idx,
 601                                                    XEN_NETIF_RSP_ERROR);
 602                 }
 603                 (*gopp_copy)++;
 604         }
 605
 606 check_frags:
 607         for (i = 0; i < nr_frags; i++, gop_map++) {
 608                 int j, newerr;
 609
 610                 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 611
 612                 /* Check error status: if okay then remember grant handle. */
 613                 newerr = gop_map->status;
 614
 615                 if (likely(!newerr)) {
 616                         xenvif_grant_handle_set(queue,
 617                                                 pending_idx,
 618                                                 gop_map->handle);
 619                         /* Had a previous error? Invalidate this fragment. */
 620                         if (unlikely(err)) {
 621                                 xenvif_idx_unmap(queue, pending_idx);
 622                                 /* If the mapping of the first frag was OK, but
 623                                  * the header's copy failed, and they are
 624                                  * sharing a slot, send an error
 625                                  */
 626                                 if (i == 0 && !first_shinfo && sharedslot)
 627                                         xenvif_idx_release(queue, pending_idx,
 628                                                            XEN_NETIF_RSP_ERROR);
 629                                 else
 630                                         xenvif_idx_release(queue, pending_idx,
 631                                                            XEN_NETIF_RSP_OKAY);
 632                         }
 633                         continue;
 634                 }
 635
 636                 /* Error on this fragment: respond to client with an error. */
 637                 if (net_ratelimit())
 638                         netdev_dbg(queue->vif->dev,
 639                                    "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 640                                    i,
 641                                    gop_map->status,
 642                                    pending_idx,
 643                                    gop_map->ref);
 644
 645                 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 646
 647                 /* Not the first error? Preceding frags already invalidated. */
 648                 if (err)
 649                         continue;
 650
 651                 /* Invalidate preceding fragments of this skb. */
 652                 for (j = 0; j < i; j++) {
 653                         pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 654                         xenvif_idx_unmap(queue, pending_idx);
 655                         xenvif_idx_release(queue, pending_idx,
 656                                            XEN_NETIF_RSP_OKAY);
 657                 }
 658
 659                 /* And if we found the error while checking the frag_list, unmap
 660                  * the first skb's frags
 661                  */
 662                 if (first_shinfo) {
 663                         for (j = 0; j < first_shinfo->nr_frags; j++) {
 664                                 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 665                                 xenvif_idx_unmap(queue, pending_idx);
 666                                 xenvif_idx_release(queue, pending_idx,
 667                                                    XEN_NETIF_RSP_OKAY);
 668                         }
 669                 }
 670
 671                 /* Remember the error: invalidate all subsequent fragments. */
 672                 err = newerr;
 673         }
 674
 675         if (skb_has_frag_list(skb) && !first_shinfo) {
 676                 first_shinfo = shinfo;
 677                 shinfo = skb_shinfo(shinfo->frag_list);
 678                 nr_frags = shinfo->nr_frags;
 679
 680                 goto check_frags;
 681         }
 682
 683         *gopp_map = gop_map;
 684         return err;
 685 }
 686
 687 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 688 {
 689         struct skb_shared_info *shinfo = skb_shinfo(skb);
 690         int nr_frags = shinfo->nr_frags;
 691         int i;
 692         u16 prev_pending_idx = INVALID_PENDING_IDX;
 693
 694         for (i = 0; i < nr_frags; i++) {
 695                 skb_frag_t *frag = shinfo->frags + i;
 696                 struct xen_netif_tx_request *txp;
 697                 struct page *page;
 698                 u16 pending_idx;
 699
 700                 pending_idx = frag_get_pending_idx(frag);
 701
 702                 /* If this is not the first frag, chain it to the previous*/
 703                 if (prev_pending_idx == INVALID_PENDING_IDX)
 704                         skb_shinfo(skb)->destructor_arg =
 705                                 &callback_param(queue, pending_idx);
 706                 else
 707                         callback_param(queue, prev_pending_idx).ctx =
 708                                 &callback_param(queue, pending_idx);
 709
 710                 callback_param(queue, pending_idx).ctx = NULL;
 711                 prev_pending_idx = pending_idx;
 712
 713                 txp = &queue->pending_tx_info[pending_idx].req;
 714                 page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
 715                 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 716                 skb->len += txp->size;
 717                 skb->data_len += txp->size;
 718                 skb->truesize += txp->size;
 719
 720                 /* Take an extra reference to offset network stack's put_page */
 721                 get_page(queue->mmap_pages[pending_idx]);
 722         }
 723 }
 724
 725 static int xenvif_get_extras(struct xenvif_queue *queue,
 726                              struct xen_netif_extra_info *extras,
 727                              unsigned int *extra_count,
 728                              int work_to_do)
 729 {
 730         struct xen_netif_extra_info extra;
 731         RING_IDX cons = queue->tx.req_cons;
 732
 733         do {
 734                 if (unlikely(work_to_do-- <= 0)) {
 735                         netdev_err(queue->vif->dev, "Missing extra info\n");
 736                         xenvif_fatal_tx_err(queue->vif);
 737                         return -EBADR;
 738                 }
 739
 740                 RING_COPY_REQUEST(&queue->tx, cons, &extra);
 741
 742                 queue->tx.req_cons = ++cons;
 743                 (*extra_count)++;
 744
 745                 if (unlikely(!extra.type ||
 746                              extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 747                         netdev_err(queue->vif->dev,
 748                                    "Invalid extra type: %d\n", extra.type);
 749                         xenvif_fatal_tx_err(queue->vif);
 750                         return -EINVAL;
 751                 }
 752
 753                 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 754         } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 755
 756         return work_to_do;
 757 }
 758
 759 static int xenvif_set_skb_gso(struct xenvif *vif,
 760                               struct sk_buff *skb,
 761                               struct xen_netif_extra_info *gso)
 762 {
 763         if (!gso->u.gso.size) {
 764                 netdev_err(vif->dev, "GSO size must not be zero.\n");
 765                 xenvif_fatal_tx_err(vif);
 766                 return -EINVAL;
 767         }
 768
 769         switch (gso->u.gso.type) {
 770         case XEN_NETIF_GSO_TYPE_TCPV4:
 771                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 772                 break;
 773         case XEN_NETIF_GSO_TYPE_TCPV6:
 774                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 775                 break;
 776         default:
 777                 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 778                 xenvif_fatal_tx_err(vif);
 779                 return -EINVAL;
 780         }
 781
 782         skb_shinfo(skb)->gso_size = gso->u.gso.size;
 783         /* gso_segs will be calculated later */
 784
 785         return 0;
 786 }
 787
 788 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 789 {
 790         bool recalculate_partial_csum = false;
 791
 792         /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 793          * peers can fail to set NETRXF_csum_blank when sending a GSO
 794          * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 795          * recalculate the partial checksum.
 796          */
 797         if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 798                 queue->stats.rx_gso_checksum_fixup++;
 799                 skb->ip_summed = CHECKSUM_PARTIAL;
 800                 recalculate_partial_csum = true;
 801         }
 802
 803         /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 804         if (skb->ip_summed != CHECKSUM_PARTIAL)
 805                 return 0;
 806
 807         return skb_checksum_setup(skb, recalculate_partial_csum);
 808 }
 809
 810 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 811 {
 812         u64 now = get_jiffies_64();
 813         u64 next_credit = queue->credit_window_start +
 814                 msecs_to_jiffies(queue->credit_usec / 1000);
 815
 816         /* Timer could already be pending in rare cases. */
 817         if (timer_pending(&queue->credit_timeout)) {
 818                 queue->rate_limited = true;
 819                 return true;
 820         }
 821
 822         /* Passed the point where we can replenish credit? */
 823         if (time_after_eq64(now, next_credit)) {
 824                 queue->credit_window_start = now;
 825                 tx_add_credit(queue);
 826         }
 827
 828         /* Still too big to send right now? Set a callback. */
 829         if (size > queue->remaining_credit) {
 830                 mod_timer(&queue->credit_timeout,
 831                           next_credit);
 832                 queue->credit_window_start = next_credit;
 833                 queue->rate_limited = true;
 834
 835                 return true;
 836         }
 837
 838         return false;
 839 }
 840
 841 /* No locking is required in xenvif_mcast_add/del() as they are
 842  * only ever invoked from NAPI poll. An RCU list is used because
 843  * xenvif_mcast_match() is called asynchronously, during start_xmit.
 844  */
 845
 846 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 847 {
 848         struct xenvif_mcast_addr *mcast;
 849
 850         if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 851                 if (net_ratelimit())
 852                         netdev_err(vif->dev,
 853                                    "Too many multicast addresses\n");
 854                 return -ENOSPC;
 855         }
 856
 857         mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 858         if (!mcast)
 859                 return -ENOMEM;
 860
 861         ether_addr_copy(mcast->addr, addr);
 862         list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 863         vif->fe_mcast_count++;
 864
 865         return 0;
 866 }
 867
 868 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 869 {
 870         struct xenvif_mcast_addr *mcast;
 871
 872         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 873                 if (ether_addr_equal(addr, mcast->addr)) {
 874                         --vif->fe_mcast_count;
 875                         list_del_rcu(&mcast->entry);
 876                         kfree_rcu(mcast, rcu);
 877                         break;
 878                 }
 879         }
 880 }
 881
 882 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 883 {
 884         struct xenvif_mcast_addr *mcast;
 885
 886         rcu_read_lock();
 887         list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 888                 if (ether_addr_equal(addr, mcast->addr)) {
 889                         rcu_read_unlock();
 890                         return true;
 891                 }
 892         }
 893         rcu_read_unlock();
 894
 895         return false;
 896 }
 897
 898 void xenvif_mcast_addr_list_free(struct xenvif *vif)
 899 {
 900         /* No need for locking or RCU here. NAPI poll and TX queue
 901          * are stopped.
 902          */
 903         while (!list_empty(&vif->fe_mcast_addr)) {
 904                 struct xenvif_mcast_addr *mcast;
 905
 906                 mcast = list_first_entry(&vif->fe_mcast_addr,
 907                                          struct xenvif_mcast_addr,
 908                                          entry);
 909                 --vif->fe_mcast_count;
 910                 list_del(&mcast->entry);
 911                 kfree(mcast);
 912         }
 913 }
 914
 915 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 916                                      int budget,
 917                                      unsigned *copy_ops,
 918                                      unsigned *map_ops)
 919 {
 920         struct sk_buff *skb, *nskb;
 921         int ret;
 922         unsigned int frag_overflow;
 923
 924         while (skb_queue_len(&queue->tx_queue) < budget) {
 925                 struct xen_netif_tx_request txreq;
 926                 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 927                 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 928                 unsigned int extra_count;
 929                 RING_IDX idx;
 930                 int work_to_do;
 931                 unsigned int data_len;
 932
 933                 if (queue->tx.sring->req_prod - queue->tx.req_cons >
 934                     XEN_NETIF_TX_RING_SIZE) {
 935                         netdev_err(queue->vif->dev,
 936                                    "Impossible number of requests. "
 937                                    "req_prod %d, req_cons %d, size %ld\n",
 938                                    queue->tx.sring->req_prod, queue->tx.req_cons,
 939                                    XEN_NETIF_TX_RING_SIZE);
 940                         xenvif_fatal_tx_err(queue->vif);
 941                         break;
 942                 }
 943
 944                 work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
 945                 if (!work_to_do)
 946                         break;
 947
 948                 idx = queue->tx.req_cons;
 949                 rmb(); /* Ensure that we see the request before we copy it. */
 950                 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 951
 952                 /* Credit-based scheduling. */
 953                 if (txreq.size > queue->remaining_credit &&
 954                     tx_credit_exceeded(queue, txreq.size))
 955                         break;
 956
 957                 queue->remaining_credit -= txreq.size;
 958
 959                 work_to_do--;
 960                 queue->tx.req_cons = ++idx;
 961
 962                 memset(extras, 0, sizeof(extras));
 963                 extra_count = 0;
 964                 if (txreq.flags & XEN_NETTXF_extra_info) {
 965                         work_to_do = xenvif_get_extras(queue, extras,
 966                                                        &extra_count,
 967                                                        work_to_do);
 968                         idx = queue->tx.req_cons;
 969                         if (unlikely(work_to_do < 0))
 970                                 break;
 971                 }
 972
 973                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 974                         struct xen_netif_extra_info *extra;
 975
 976                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 977                         ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 978
 979                         make_tx_response(queue, &txreq, extra_count,
 980                                          (ret == 0) ?
 981                                          XEN_NETIF_RSP_OKAY :
 982                                          XEN_NETIF_RSP_ERROR);
 983                         continue;
 984                 }
 985
 986                 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 987                         struct xen_netif_extra_info *extra;
 988
 989                         extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 990                         xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 991
 992                         make_tx_response(queue, &txreq, extra_count,
 993                                          XEN_NETIF_RSP_OKAY);
 994                         continue;
 995                 }
 996
 997                 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
 998                         XEN_NETBACK_TX_COPY_LEN : txreq.size;
 999
1000                 ret = xenvif_count_requests(queue, &txreq, extra_count,
1001                                             txfrags, work_to_do);
1002
1003                 if (unlikely(ret < 0))
1004                         break;
1005
1006                 idx += ret;
1007
1008                 if (unlikely(txreq.size < ETH_HLEN)) {
1009                         netdev_dbg(queue->vif->dev,
1010                                    "Bad packet size: %d\n", txreq.size);
1011                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1012                         break;
1013                 }
1014
1015                 /* No crossing a page as the payload mustn't fragment. */
1016                 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1017                         netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1018                                    txreq.offset, txreq.size);
1019                         xenvif_fatal_tx_err(queue->vif);
1020                         break;
1021                 }
1022
1023                 if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1024                         data_len = txreq.size;
1025
1026                 skb = xenvif_alloc_skb(data_len);
1027                 if (unlikely(skb == NULL)) {
1028                         netdev_dbg(queue->vif->dev,
1029                                    "Can't allocate a skb in start_xmit.\n");
1030                         xenvif_tx_err(queue, &txreq, extra_count, idx);
1031                         break;
1032                 }
1033
1034                 skb_shinfo(skb)->nr_frags = ret;
1035                 /* At this point shinfo->nr_frags is in fact the number of
1036                  * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1037                  */
1038                 frag_overflow = 0;
1039                 nskb = NULL;
1040                 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1041                         frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1042                         BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1043                         skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1044                         nskb = xenvif_alloc_skb(0);
1045                         if (unlikely(nskb == NULL)) {
1046                                 skb_shinfo(skb)->nr_frags = 0;
1047                                 kfree_skb(skb);
1048                                 xenvif_tx_err(queue, &txreq, extra_count, idx);
1049                                 if (net_ratelimit())
1050                                         netdev_err(queue->vif->dev,
1051                                                    "Can't allocate the frag_list skb.\n");
1052                                 break;
1053                         }
1054                 }
1055
1056                 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1057                         struct xen_netif_extra_info *gso;
1058                         gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1059
1060                         if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1061                                 /* Failure in xenvif_set_skb_gso is fatal. */
1062                                 skb_shinfo(skb)->nr_frags = 0;
1063                                 kfree_skb(skb);
1064                                 kfree_skb(nskb);
1065                                 break;
1066                         }
1067                 }
1068
1069                 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1070                         struct xen_netif_extra_info *extra;
1071                         enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1072
1073                         extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1074
1075                         switch (extra->u.hash.type) {
1076                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1077                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1078                                 type = PKT_HASH_TYPE_L3;
1079                                 break;
1080
1081                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1082                         case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1083                                 type = PKT_HASH_TYPE_L4;
1084                                 break;
1085
1086                         default:
1087                                 break;
1088                         }
1089
1090                         if (type != PKT_HASH_TYPE_NONE)
1091                                 skb_set_hash(skb,
1092                                              *(u32 *)extra->u.hash.value,
1093                                              type);
1094                 }
1095
1096                 xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1097                                     map_ops, frag_overflow, nskb, extra_count,
1098                                     data_len);
1099
1100                 __skb_queue_tail(&queue->tx_queue, skb);
1101
1102                 queue->tx.req_cons = idx;
1103         }
1104
1105         return;
1106 }
1107
1108 /* Consolidate skb with a frag_list into a brand new one with local pages on
1109  * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1110  */
1111 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1112 {
1113         unsigned int offset = skb_headlen(skb);
1114         skb_frag_t frags[MAX_SKB_FRAGS];
1115         int i, f;
1116         struct ubuf_info *uarg;
1117         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1118
1119         queue->stats.tx_zerocopy_sent += 2;
1120         queue->stats.tx_frag_overflow++;
1121
1122         xenvif_fill_frags(queue, nskb);
1123         /* Subtract frags size, we will correct it later */
1124         skb->truesize -= skb->data_len;
1125         skb->len += nskb->len;
1126         skb->data_len += nskb->len;
1127
1128         /* create a brand new frags array and coalesce there */
1129         for (i = 0; offset < skb->len; i++) {
1130                 struct page *page;
1131                 unsigned int len;
1132
1133                 BUG_ON(i >= MAX_SKB_FRAGS);
1134                 page = alloc_page(GFP_ATOMIC);
1135                 if (!page) {
1136                         int j;
1137                         skb->truesize += skb->data_len;
1138                         for (j = 0; j < i; j++)
1139                                 put_page(skb_frag_page(&frags[j]));
1140                         return -ENOMEM;
1141                 }
1142
1143                 if (offset + PAGE_SIZE < skb->len)
1144                         len = PAGE_SIZE;
1145                 else
1146                         len = skb->len - offset;
1147                 if (skb_copy_bits(skb, offset, page_address(page), len))
1148                         BUG();
1149
1150                 offset += len;
1151                 skb_frag_fill_page_desc(&frags[i], page, 0, len);
1152         }
1153
1154         /* Release all the original (foreign) frags. */
1155         for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1156                 skb_frag_unref(skb, f);
1157         uarg = skb_shinfo(skb)->destructor_arg;
1158         /* increase inflight counter to offset decrement in callback */
1159         atomic_inc(&queue->inflight_packets);
1160         uarg->ops->complete(NULL, uarg, true);
1161         skb_shinfo(skb)->destructor_arg = NULL;
1162
1163         /* Fill the skb with the new (local) frags. */
1164         memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1165         skb_shinfo(skb)->nr_frags = i;
1166         skb->truesize += i * PAGE_SIZE;
1167
1168         return 0;
1169 }
1170
1171 static int xenvif_tx_submit(struct xenvif_queue *queue)
1172 {
1173         struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1174         struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1175         struct sk_buff *skb;
1176         int work_done = 0;
1177
1178         while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1179                 struct xen_netif_tx_request *txp;
1180                 u16 pending_idx;
1181
1182                 pending_idx = copy_pending_idx(skb, 0);
1183                 txp = &queue->pending_tx_info[pending_idx].req;
1184
1185                 /* Check the remap error code. */
1186                 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1187                         /* If there was an error, xenvif_tx_check_gop is
1188                          * expected to release all the frags which were mapped,
1189                          * so kfree_skb shouldn't do it again
1190                          */
1191                         skb_shinfo(skb)->nr_frags = 0;
1192                         if (skb_has_frag_list(skb)) {
1193                                 struct sk_buff *nskb =
1194                                                 skb_shinfo(skb)->frag_list;
1195                                 skb_shinfo(nskb)->nr_frags = 0;
1196                         }
1197                         kfree_skb(skb);
1198                         continue;
1199                 }
1200
1201                 if (txp->flags & XEN_NETTXF_csum_blank)
1202                         skb->ip_summed = CHECKSUM_PARTIAL;
1203                 else if (txp->flags & XEN_NETTXF_data_validated)
1204                         skb->ip_summed = CHECKSUM_UNNECESSARY;
1205
1206                 xenvif_fill_frags(queue, skb);
1207
1208                 if (unlikely(skb_has_frag_list(skb))) {
1209                         struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1210                         xenvif_skb_zerocopy_prepare(queue, nskb);
1211                         if (xenvif_handle_frag_list(queue, skb)) {
1212                                 if (net_ratelimit())
1213                                         netdev_err(queue->vif->dev,
1214                                                    "Not enough memory to consolidate frag_list!\n");
1215                                 xenvif_skb_zerocopy_prepare(queue, skb);
1216                                 kfree_skb(skb);
1217                                 continue;
1218                         }
1219                         /* Copied all the bits from the frag list -- free it. */
1220                         skb_frag_list_init(skb);
1221                         kfree_skb(nskb);
1222                 }
1223
1224                 skb->dev      = queue->vif->dev;
1225                 skb->protocol = eth_type_trans(skb, skb->dev);
1226                 skb_reset_network_header(skb);
1227
1228                 if (checksum_setup(queue, skb)) {
1229                         netdev_dbg(queue->vif->dev,
1230                                    "Can't setup checksum in net_tx_action\n");
1231                         /* We have to set this flag to trigger the callback */
1232                         if (skb_shinfo(skb)->destructor_arg)
1233                                 xenvif_skb_zerocopy_prepare(queue, skb);
1234                         kfree_skb(skb);
1235                         continue;
1236                 }
1237
1238                 skb_probe_transport_header(skb);
1239
1240                 /* If the packet is GSO then we will have just set up the
1241                  * transport header offset in checksum_setup so it's now
1242                  * straightforward to calculate gso_segs.
1243                  */
1244                 if (skb_is_gso(skb)) {
1245                         int mss, hdrlen;
1246
1247                         /* GSO implies having the L4 header. */
1248                         WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1249                         if (unlikely(!skb_transport_header_was_set(skb))) {
1250                                 kfree_skb(skb);
1251                                 continue;
1252                         }
1253
1254                         mss = skb_shinfo(skb)->gso_size;
1255                         hdrlen = skb_tcp_all_headers(skb);
1256
1257                         skb_shinfo(skb)->gso_segs =
1258                                 DIV_ROUND_UP(skb->len - hdrlen, mss);
1259                 }
1260
1261                 queue->stats.rx_bytes += skb->len;
1262                 queue->stats.rx_packets++;
1263
1264                 work_done++;
1265
1266                 /* Set this flag right before netif_receive_skb, otherwise
1267                  * someone might think this packet already left netback, and
1268                  * do a skb_copy_ubufs while we are still in control of the
1269                  * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1270                  */
1271                 if (skb_shinfo(skb)->destructor_arg) {
1272                         xenvif_skb_zerocopy_prepare(queue, skb);
1273                         queue->stats.tx_zerocopy_sent++;
1274                 }
1275
1276                 netif_receive_skb(skb);
1277         }
1278
1279         return work_done;
1280 }
1281
1282 static void xenvif_zerocopy_callback(struct sk_buff *skb,
1283                                      struct ubuf_info *ubuf_base,
1284                                      bool zerocopy_success)
1285 {
1286         unsigned long flags;
1287         pending_ring_idx_t index;
1288         struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1289         struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1290
1291         /* This is the only place where we grab this lock, to protect callbacks
1292          * from each other.
1293          */
1294         spin_lock_irqsave(&queue->callback_lock, flags);
1295         do {
1296                 u16 pending_idx = ubuf->desc;
1297                 ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1298                 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1299                         MAX_PENDING_REQS);
1300                 index = pending_index(queue->dealloc_prod);
1301                 queue->dealloc_ring[index] = pending_idx;
1302                 /* Sync with xenvif_tx_dealloc_action:
1303                  * insert idx then incr producer.
1304                  */
1305                 smp_wmb();
1306                 queue->dealloc_prod++;
1307         } while (ubuf);
1308         spin_unlock_irqrestore(&queue->callback_lock, flags);
1309
1310         if (likely(zerocopy_success))
1311                 queue->stats.tx_zerocopy_success++;
1312         else
1313                 queue->stats.tx_zerocopy_fail++;
1314         xenvif_skb_zerocopy_complete(queue);
1315 }
1316
1317 const struct ubuf_info_ops xenvif_ubuf_ops = {
1318         .complete = xenvif_zerocopy_callback,
1319 };
1320
1321 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1322 {
1323         struct gnttab_unmap_grant_ref *gop;
1324         pending_ring_idx_t dc, dp;
1325         u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1326         unsigned int i = 0;
1327
1328         dc = queue->dealloc_cons;
1329         gop = queue->tx_unmap_ops;
1330
1331         /* Free up any grants we have finished using */
1332         do {
1333                 dp = queue->dealloc_prod;
1334
1335                 /* Ensure we see all indices enqueued by all
1336                  * xenvif_zerocopy_callback().
1337                  */
1338                 smp_rmb();
1339
1340                 while (dc != dp) {
1341                         BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1342                         pending_idx =
1343                                 queue->dealloc_ring[pending_index(dc++)];
1344
1345                         pending_idx_release[gop - queue->tx_unmap_ops] =
1346                                 pending_idx;
1347                         queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1348                                 queue->mmap_pages[pending_idx];
1349                         gnttab_set_unmap_op(gop,
1350                                             idx_to_kaddr(queue, pending_idx),
1351                                             GNTMAP_host_map,
1352                                             queue->grant_tx_handle[pending_idx]);
1353                         xenvif_grant_handle_reset(queue, pending_idx);
1354                         ++gop;
1355                 }
1356
1357         } while (dp != queue->dealloc_prod);
1358
1359         queue->dealloc_cons = dc;
1360
1361         if (gop - queue->tx_unmap_ops > 0) {
1362                 int ret;
1363                 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1364                                         NULL,
1365                                         queue->pages_to_unmap,
1366                                         gop - queue->tx_unmap_ops);
1367                 if (ret) {
1368                         netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1369                                    gop - queue->tx_unmap_ops, ret);
1370                         for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1371                                 if (gop[i].status != GNTST_okay)
1372                                         netdev_err(queue->vif->dev,
1373                                                    " host_addr: 0x%llx handle: 0x%x status: %d\n",
1374                                                    gop[i].host_addr,
1375                                                    gop[i].handle,
1376                                                    gop[i].status);
1377                         }
1378                         BUG();
1379                 }
1380         }
1381
1382         for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1383                 xenvif_idx_release(queue, pending_idx_release[i],
1384                                    XEN_NETIF_RSP_OKAY);
1385 }
1386
1387
1388 /* Called after netfront has transmitted */
1389 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1390 {
1391         unsigned nr_mops = 0, nr_cops = 0;
1392         int work_done, ret;
1393
1394         if (unlikely(!tx_work_todo(queue)))
1395                 return 0;
1396
1397         xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1398
1399         if (nr_cops == 0)
1400                 return 0;
1401
1402         gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1403         if (nr_mops != 0) {
1404                 ret = gnttab_map_refs(queue->tx_map_ops,
1405                                       NULL,
1406                                       queue->pages_to_map,
1407                                       nr_mops);
1408                 if (ret) {
1409                         unsigned int i;
1410
1411                         netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1412                                    nr_mops, ret);
1413                         for (i = 0; i < nr_mops; ++i)
1414                                 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1415                                              GNTST_okay);
1416                 }
1417         }
1418
1419         work_done = xenvif_tx_submit(queue);
1420
1421         return work_done;
1422 }
1423
1424 static void _make_tx_response(struct xenvif_queue *queue,
1425                              const struct xen_netif_tx_request *txp,
1426                              unsigned int extra_count,
1427                              s8 status)
1428 {
1429         RING_IDX i = queue->tx.rsp_prod_pvt;
1430         struct xen_netif_tx_response *resp;
1431
1432         resp = RING_GET_RESPONSE(&queue->tx, i);
1433         resp->id     = txp->id;
1434         resp->status = status;
1435
1436         while (extra_count-- != 0)
1437                 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1438
1439         queue->tx.rsp_prod_pvt = ++i;
1440 }
1441
1442 static void push_tx_responses(struct xenvif_queue *queue)
1443 {
1444         int notify;
1445
1446         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1447         if (notify)
1448                 notify_remote_via_irq(queue->tx_irq);
1449 }
1450
1451 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1452                                s8 status)
1453 {
1454         struct pending_tx_info *pending_tx_info;
1455         pending_ring_idx_t index;
1456         unsigned long flags;
1457
1458         pending_tx_info = &queue->pending_tx_info[pending_idx];
1459
1460         spin_lock_irqsave(&queue->response_lock, flags);
1461
1462         _make_tx_response(queue, &pending_tx_info->req,
1463                           pending_tx_info->extra_count, status);
1464
1465         /* Release the pending index before pusing the Tx response so
1466          * its available before a new Tx request is pushed by the
1467          * frontend.
1468          */
1469         index = pending_index(queue->pending_prod++);
1470         queue->pending_ring[index] = pending_idx;
1471
1472         push_tx_responses(queue);
1473
1474         spin_unlock_irqrestore(&queue->response_lock, flags);
1475 }
1476
1477 static void make_tx_response(struct xenvif_queue *queue,
1478                              const struct xen_netif_tx_request *txp,
1479                              unsigned int extra_count,
1480                              s8 status)
1481 {
1482         unsigned long flags;
1483
1484         spin_lock_irqsave(&queue->response_lock, flags);
1485
1486         _make_tx_response(queue, txp, extra_count, status);
1487         push_tx_responses(queue);
1488
1489         spin_unlock_irqrestore(&queue->response_lock, flags);
1490 }
1491
1492 static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1493 {
1494         int ret;
1495         struct gnttab_unmap_grant_ref tx_unmap_op;
1496
1497         gnttab_set_unmap_op(&tx_unmap_op,
1498                             idx_to_kaddr(queue, pending_idx),
1499                             GNTMAP_host_map,
1500                             queue->grant_tx_handle[pending_idx]);
1501         xenvif_grant_handle_reset(queue, pending_idx);
1502
1503         ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1504                                 &queue->mmap_pages[pending_idx], 1);
1505         if (ret) {
1506                 netdev_err(queue->vif->dev,
1507                            "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1508                            ret,
1509                            pending_idx,
1510                            tx_unmap_op.host_addr,
1511                            tx_unmap_op.handle,
1512                            tx_unmap_op.status);
1513                 BUG();
1514         }
1515 }
1516
1517 static inline int tx_work_todo(struct xenvif_queue *queue)
1518 {
1519         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1520                 return 1;
1521
1522         return 0;
1523 }
1524
1525 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1526 {
1527         return queue->dealloc_cons != queue->dealloc_prod;
1528 }
1529
1530 void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1531 {
1532         if (queue->tx.sring)
1533                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1534                                         queue->tx.sring);
1535         if (queue->rx.sring)
1536                 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1537                                         queue->rx.sring);
1538 }
1539
1540 int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1541                                    grant_ref_t tx_ring_ref,
1542                                    grant_ref_t rx_ring_ref)
1543 {
1544         void *addr;
1545         struct xen_netif_tx_sring *txs;
1546         struct xen_netif_rx_sring *rxs;
1547         RING_IDX rsp_prod, req_prod;
1548         int err;
1549
1550         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1551                                      &tx_ring_ref, 1, &addr);
1552         if (err)
1553                 goto err;
1554
1555         txs = (struct xen_netif_tx_sring *)addr;
1556         rsp_prod = READ_ONCE(txs->rsp_prod);
1557         req_prod = READ_ONCE(txs->req_prod);
1558
1559         BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1560
1561         err = -EIO;
1562         if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1563                 goto err;
1564
1565         err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1566                                      &rx_ring_ref, 1, &addr);
1567         if (err)
1568                 goto err;
1569
1570         rxs = (struct xen_netif_rx_sring *)addr;
1571         rsp_prod = READ_ONCE(rxs->rsp_prod);
1572         req_prod = READ_ONCE(rxs->req_prod);
1573
1574         BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1575
1576         err = -EIO;
1577         if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1578                 goto err;
1579
1580         return 0;
1581
1582 err:
1583         xenvif_unmap_frontend_data_rings(queue);
1584         return err;
1585 }
1586
1587 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1588 {
1589         /* Dealloc thread must remain running until all inflight
1590          * packets complete.
1591          */
1592         return kthread_should_stop() &&
1593                 !atomic_read(&queue->inflight_packets);
1594 }
1595
1596 int xenvif_dealloc_kthread(void *data)
1597 {
1598         struct xenvif_queue *queue = data;
1599
1600         for (;;) {
1601                 wait_event_interruptible(queue->dealloc_wq,
1602                                          tx_dealloc_work_todo(queue) ||
1603                                          xenvif_dealloc_kthread_should_stop(queue));
1604                 if (xenvif_dealloc_kthread_should_stop(queue))
1605                         break;
1606
1607                 xenvif_tx_dealloc_action(queue);
1608                 cond_resched();
1609         }
1610
1611         /* Unmap anything remaining*/
1612         if (tx_dealloc_work_todo(queue))
1613                 xenvif_tx_dealloc_action(queue);
1614
1615         return 0;
1616 }
1617
1618 static void make_ctrl_response(struct xenvif *vif,
1619                                const struct xen_netif_ctrl_request *req,
1620                                u32 status, u32 data)
1621 {
1622         RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1623         struct xen_netif_ctrl_response rsp = {
1624                 .id = req->id,
1625                 .type = req->type,
1626                 .status = status,
1627                 .data = data,
1628         };
1629
1630         *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1631         vif->ctrl.rsp_prod_pvt = ++idx;
1632 }
1633
1634 static void push_ctrl_response(struct xenvif *vif)
1635 {
1636         int notify;
1637
1638         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1639         if (notify)
1640                 notify_remote_via_irq(vif->ctrl_irq);
1641 }
1642
1643 static void process_ctrl_request(struct xenvif *vif,
1644                                  const struct xen_netif_ctrl_request *req)
1645 {
1646         u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1647         u32 data = 0;
1648
1649         switch (req->type) {
1650         case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1651                 status = xenvif_set_hash_alg(vif, req->data[0]);
1652                 break;
1653
1654         case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1655                 status = xenvif_get_hash_flags(vif, &data);
1656                 break;
1657
1658         case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1659                 status = xenvif_set_hash_flags(vif, req->data[0]);
1660                 break;
1661
1662         case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1663                 status = xenvif_set_hash_key(vif, req->data[0],
1664                                              req->data[1]);
1665                 break;
1666
1667         case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1668                 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1669                 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1670                 break;
1671
1672         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1673                 status = xenvif_set_hash_mapping_size(vif,
1674                                                       req->data[0]);
1675                 break;
1676
1677         case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1678                 status = xenvif_set_hash_mapping(vif, req->data[0],
1679                                                  req->data[1],
1680                                                  req->data[2]);
1681                 break;
1682
1683         default:
1684                 break;
1685         }
1686
1687         make_ctrl_response(vif, req, status, data);
1688         push_ctrl_response(vif);
1689 }
1690
1691 static void xenvif_ctrl_action(struct xenvif *vif)
1692 {
1693         for (;;) {
1694                 RING_IDX req_prod, req_cons;
1695
1696                 req_prod = vif->ctrl.sring->req_prod;
1697                 req_cons = vif->ctrl.req_cons;
1698
1699                 /* Make sure we can see requests before we process them. */
1700                 rmb();
1701
1702                 if (req_cons == req_prod)
1703                         break;
1704
1705                 while (req_cons != req_prod) {
1706                         struct xen_netif_ctrl_request req;
1707
1708                         RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1709                         req_cons++;
1710
1711                         process_ctrl_request(vif, &req);
1712                 }
1713
1714                 vif->ctrl.req_cons = req_cons;
1715                 vif->ctrl.sring->req_event = req_cons + 1;
1716         }
1717 }
1718
1719 static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1720 {
1721         if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1722                 return true;
1723
1724         return false;
1725 }
1726
1727 irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1728 {
1729         struct xenvif *vif = data;
1730         unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1731
1732         while (xenvif_ctrl_work_todo(vif)) {
1733                 xenvif_ctrl_action(vif);
1734                 eoi_flag = 0;
1735         }
1736
1737         xen_irq_lateeoi(irq, eoi_flag);
1738
1739         return IRQ_HANDLED;
1740 }
1741
1742 static int __init netback_init(void)
1743 {
1744         int rc = 0;
1745
1746         if (!xen_domain())
1747                 return -ENODEV;
1748
1749         /* Allow as many queues as there are CPUs but max. 8 if user has not
1750          * specified a value.
1751          */
1752         if (xenvif_max_queues == 0)
1753                 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1754                                           num_online_cpus());
1755
1756         if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1757                 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1758                         fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1759                 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1760         }
1761
1762         rc = xenvif_xenbus_init();
1763         if (rc)
1764                 goto failed_init;
1765
1766 #ifdef CONFIG_DEBUG_FS
1767         xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1768 #endif /* CONFIG_DEBUG_FS */
1769
1770         return 0;
1771
1772 failed_init:
1773         return rc;
1774 }
1775
1776 module_init(netback_init);
1777
1778 static void __exit netback_fini(void)
1779 {
1780 #ifdef CONFIG_DEBUG_FS
1781         debugfs_remove_recursive(xen_netback_dbg_root);
1782 #endif /* CONFIG_DEBUG_FS */
1783         xenvif_xenbus_fini();
1784 }
1785 module_exit(netback_fini);
1786
1787 MODULE_DESCRIPTION("Xen backend network device module");
1788 MODULE_LICENSE("Dual BSD/GPL");
1789 MODULE_ALIAS("xen-backend:vif");